2 * PPC Huge TLB Page Support for Kernel.
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
5 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
7 * Based on the IA-32 version:
8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
13 #include <linux/slab.h>
14 #include <linux/hugetlb.h>
15 #include <linux/export.h>
16 #include <linux/of_fdt.h>
17 #include <linux/memblock.h>
18 #include <linux/moduleparam.h>
19 #include <linux/swap.h>
20 #include <linux/swapops.h>
21 #include <linux/kmemleak.h>
22 #include <asm/pgalloc.h>
24 #include <asm/setup.h>
25 #include <asm/hugetlb.h>
26 #include <asm/pte-walk.h>
27 #include <asm/firmware.h>
29 bool hugetlb_disabled = false;
31 #define hugepd_none(hpd) (hpd_val(hpd) == 0)
33 #define PTE_T_ORDER (__builtin_ffs(sizeof(pte_basic_t)) - \
34 __builtin_ffs(sizeof(void *)))
36 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
39 * Only called for hugetlbfs pages, hence can ignore THP and the
42 return __find_linux_pte(mm->pgd, addr, NULL, NULL);
45 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
46 unsigned long address, unsigned int pdshift,
47 unsigned int pshift, spinlock_t *ptl)
49 struct kmem_cache *cachep;
54 if (pshift >= pdshift) {
55 cachep = PGT_CACHE(PTE_T_ORDER);
56 num_hugepd = 1 << (pshift - pdshift);
58 cachep = PGT_CACHE(pdshift - pshift);
63 WARN_ONCE(1, "No page table cache created for hugetlb tables");
67 new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
69 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
70 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
76 * Make sure other cpus find the hugepd set only after a
77 * properly initialized page table is visible to them.
78 * For more details look for comment in __pte_alloc().
84 * We have multiple higher-level entries that point to the same
85 * actual pte location. Fill in each as we go and backtrack on error.
86 * We need all of these so the DTLB pgtable walk code can find the
87 * right higher-level entry without knowing if it's a hugepage or not.
89 for (i = 0; i < num_hugepd; i++, hpdp++) {
90 if (unlikely(!hugepd_none(*hpdp)))
92 hugepd_populate(hpdp, new, pshift);
94 /* If we bailed from the for loop early, an error occurred, clean up */
96 for (i = i - 1 ; i >= 0; i--, hpdp--)
98 kmem_cache_free(cachep, new);
100 kmemleak_ignore(new);
107 * At this point we do the placement change only for BOOK3S 64. This would
108 * possibly work on other subarchs.
110 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
111 unsigned long addr, unsigned long sz)
117 hugepd_t *hpdp = NULL;
118 unsigned pshift = __ffs(sz);
119 unsigned pdshift = PGDIR_SHIFT;
123 pg = pgd_offset(mm, addr);
124 p4 = p4d_offset(pg, addr);
126 #ifdef CONFIG_PPC_BOOK3S_64
127 if (pshift == PGDIR_SHIFT)
130 else if (pshift > PUD_SHIFT) {
132 * We need to use hugepd table
134 ptl = &mm->page_table_lock;
135 hpdp = (hugepd_t *)p4;
138 pu = pud_alloc(mm, p4, addr);
141 if (pshift == PUD_SHIFT)
143 else if (pshift > PMD_SHIFT) {
144 ptl = pud_lockptr(mm, pu);
145 hpdp = (hugepd_t *)pu;
148 pm = pmd_alloc(mm, pu, addr);
151 if (pshift == PMD_SHIFT)
155 ptl = pmd_lockptr(mm, pm);
156 hpdp = (hugepd_t *)pm;
161 if (pshift >= PGDIR_SHIFT) {
162 ptl = &mm->page_table_lock;
163 hpdp = (hugepd_t *)p4;
166 pu = pud_alloc(mm, p4, addr);
169 if (pshift >= PUD_SHIFT) {
170 ptl = pud_lockptr(mm, pu);
171 hpdp = (hugepd_t *)pu;
174 pm = pmd_alloc(mm, pu, addr);
177 ptl = pmd_lockptr(mm, pm);
178 hpdp = (hugepd_t *)pm;
185 if (IS_ENABLED(CONFIG_PPC_8xx) && pshift < PMD_SHIFT)
186 return pte_alloc_map(mm, (pmd_t *)hpdp, addr);
188 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
190 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
191 pdshift, pshift, ptl))
194 return hugepte_offset(*hpdp, addr, pdshift);
197 #ifdef CONFIG_PPC_BOOK3S_64
199 * Tracks gpages after the device tree is scanned and before the
200 * huge_boot_pages list is ready on pseries.
202 #define MAX_NUMBER_GPAGES 1024
203 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
204 __initdata static unsigned nr_gpages;
207 * Build list of addresses of gigantic pages. This function is used in early
208 * boot before the buddy allocator is setup.
210 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
214 while (number_of_pages > 0) {
215 gpage_freearray[nr_gpages] = addr;
222 static int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
224 struct huge_bootmem_page *m;
227 m = phys_to_virt(gpage_freearray[--nr_gpages]);
228 gpage_freearray[nr_gpages] = 0;
229 list_add(&m->list, &huge_boot_pages);
234 bool __init hugetlb_node_alloc_supported(void)
241 int __init alloc_bootmem_huge_page(struct hstate *h, int nid)
244 #ifdef CONFIG_PPC_BOOK3S_64
245 if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
246 return pseries_alloc_bootmem_huge_page(h);
248 return __alloc_bootmem_huge_page(h, nid);
251 #ifndef CONFIG_PPC_BOOK3S_64
252 #define HUGEPD_FREELIST_SIZE \
253 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
255 struct hugepd_freelist {
261 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
263 static void hugepd_free_rcu_callback(struct rcu_head *head)
265 struct hugepd_freelist *batch =
266 container_of(head, struct hugepd_freelist, rcu);
269 for (i = 0; i < batch->index; i++)
270 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
272 free_page((unsigned long)batch);
275 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
277 struct hugepd_freelist **batchp;
279 batchp = &get_cpu_var(hugepd_freelist_cur);
281 if (atomic_read(&tlb->mm->mm_users) < 2 ||
282 mm_is_thread_local(tlb->mm)) {
283 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
284 put_cpu_var(hugepd_freelist_cur);
288 if (*batchp == NULL) {
289 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
290 (*batchp)->index = 0;
293 (*batchp)->ptes[(*batchp)->index++] = hugepte;
294 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
295 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
298 put_cpu_var(hugepd_freelist_cur);
301 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
304 /* Return true when the entry to be freed maps more than the area being freed */
305 static bool range_is_outside_limits(unsigned long start, unsigned long end,
306 unsigned long floor, unsigned long ceiling,
309 if ((start & mask) < floor)
316 return end - 1 > ceiling - 1;
319 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
320 unsigned long start, unsigned long end,
321 unsigned long floor, unsigned long ceiling)
323 pte_t *hugepte = hugepd_page(*hpdp);
326 unsigned long pdmask = ~((1UL << pdshift) - 1);
327 unsigned int num_hugepd = 1;
328 unsigned int shift = hugepd_shift(*hpdp);
330 /* Note: On fsl the hpdp may be the first of several */
332 num_hugepd = 1 << (shift - pdshift);
334 if (range_is_outside_limits(start, end, floor, ceiling, pdmask))
337 for (i = 0; i < num_hugepd; i++, hpdp++)
340 if (shift >= pdshift)
341 hugepd_free(tlb, hugepte);
343 pgtable_free_tlb(tlb, hugepte,
344 get_hugepd_cache_index(pdshift - shift));
347 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
348 unsigned long addr, unsigned long end,
349 unsigned long floor, unsigned long ceiling)
351 pgtable_t token = pmd_pgtable(*pmd);
353 if (range_is_outside_limits(addr, end, floor, ceiling, PMD_MASK))
357 pte_free_tlb(tlb, token, addr);
358 mm_dec_nr_ptes(tlb->mm);
361 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
362 unsigned long addr, unsigned long end,
363 unsigned long floor, unsigned long ceiling)
373 pmd = pmd_offset(pud, addr);
374 next = pmd_addr_end(addr, end);
375 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
376 if (pmd_none_or_clear_bad(pmd))
380 * if it is not hugepd pointer, we should already find
383 WARN_ON(!IS_ENABLED(CONFIG_PPC_8xx));
385 hugetlb_free_pte_range(tlb, pmd, addr, end, floor, ceiling);
390 * Increment next by the size of the huge mapping since
391 * there may be more than one entry at this level for a
392 * single hugepage, but all of them point to
393 * the same kmem cache that holds the hugepte.
395 more = addr + (1UL << hugepd_shift(*(hugepd_t *)pmd));
399 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
400 addr, next, floor, ceiling);
401 } while (addr = next, addr != end);
403 if (range_is_outside_limits(start, end, floor, ceiling, PUD_MASK))
406 pmd = pmd_offset(pud, start & PUD_MASK);
408 pmd_free_tlb(tlb, pmd, start & PUD_MASK);
409 mm_dec_nr_pmds(tlb->mm);
412 static void hugetlb_free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
413 unsigned long addr, unsigned long end,
414 unsigned long floor, unsigned long ceiling)
422 pud = pud_offset(p4d, addr);
423 next = pud_addr_end(addr, end);
424 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
425 if (pud_none_or_clear_bad(pud))
427 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
432 * Increment next by the size of the huge mapping since
433 * there may be more than one entry at this level for a
434 * single hugepage, but all of them point to
435 * the same kmem cache that holds the hugepte.
437 more = addr + (1UL << hugepd_shift(*(hugepd_t *)pud));
441 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
442 addr, next, floor, ceiling);
444 } while (addr = next, addr != end);
446 if (range_is_outside_limits(start, end, floor, ceiling, PGDIR_MASK))
449 pud = pud_offset(p4d, start & PGDIR_MASK);
451 pud_free_tlb(tlb, pud, start & PGDIR_MASK);
452 mm_dec_nr_puds(tlb->mm);
456 * This function frees user-level page tables of a process.
458 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
459 unsigned long addr, unsigned long end,
460 unsigned long floor, unsigned long ceiling)
467 * Because there are a number of different possible pagetable
468 * layouts for hugepage ranges, we limit knowledge of how
469 * things should be laid out to the allocation path
470 * (huge_pte_alloc(), above). Everything else works out the
471 * structure as it goes from information in the hugepd
472 * pointers. That means that we can't here use the
473 * optimization used in the normal page free_pgd_range(), of
474 * checking whether we're actually covering a large enough
475 * range to have to do anything at the top level of the walk
476 * instead of at the bottom.
478 * To make sense of this, you should probably go read the big
479 * block comment at the top of the normal free_pgd_range(),
484 next = pgd_addr_end(addr, end);
485 pgd = pgd_offset(tlb->mm, addr);
486 p4d = p4d_offset(pgd, addr);
487 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
488 if (p4d_none_or_clear_bad(p4d))
490 hugetlb_free_pud_range(tlb, p4d, addr, next, floor, ceiling);
494 * Increment next by the size of the huge mapping since
495 * there may be more than one entry at the pgd level
496 * for a single hugepage, but all of them point to the
497 * same kmem cache that holds the hugepte.
499 more = addr + (1UL << hugepd_shift(*(hugepd_t *)pgd));
503 free_hugepd_range(tlb, (hugepd_t *)p4d, PGDIR_SHIFT,
504 addr, next, floor, ceiling);
506 } while (addr = next, addr != end);
509 struct page *follow_huge_pd(struct vm_area_struct *vma,
510 unsigned long address, hugepd_t hpd,
511 int flags, int pdshift)
515 struct page *page = NULL;
517 int shift = hugepd_shift(hpd);
518 struct mm_struct *mm = vma->vm_mm;
522 * hugepage directory entries are protected by mm->page_table_lock
523 * Use this instead of huge_pte_lockptr
525 ptl = &mm->page_table_lock;
528 ptep = hugepte_offset(hpd, address, pdshift);
529 if (pte_present(*ptep)) {
530 mask = (1UL << shift) - 1;
531 page = pte_page(*ptep);
532 page += ((address & mask) >> PAGE_SHIFT);
533 if (flags & FOLL_GET)
536 if (is_hugetlb_entry_migration(*ptep)) {
538 __migration_entry_wait(mm, ptep, ptl);
546 bool __init arch_hugetlb_valid_size(unsigned long size)
548 int shift = __ffs(size);
551 /* Check that it is a page size supported by the hardware and
552 * that it fits within pagetable and slice limits. */
553 if (size <= PAGE_SIZE || !is_power_of_2(size))
556 mmu_psize = check_and_get_huge_psize(shift);
560 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
565 static int __init add_huge_page_size(unsigned long long size)
567 int shift = __ffs(size);
569 if (!arch_hugetlb_valid_size((unsigned long)size))
572 hugetlb_add_hstate(shift - PAGE_SHIFT);
576 static int __init hugetlbpage_init(void)
578 bool configured = false;
581 if (hugetlb_disabled) {
582 pr_info("HugeTLB support is disabled!\n");
586 if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
587 !mmu_has_feature(MMU_FTR_16M_PAGE))
590 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
594 if (!mmu_psize_defs[psize].shift)
597 shift = mmu_psize_to_shift(psize);
599 #ifdef CONFIG_PPC_BOOK3S_64
600 if (shift > PGDIR_SHIFT)
602 else if (shift > PUD_SHIFT)
603 pdshift = PGDIR_SHIFT;
604 else if (shift > PMD_SHIFT)
609 if (shift < PUD_SHIFT)
611 else if (shift < PGDIR_SHIFT)
614 pdshift = PGDIR_SHIFT;
617 if (add_huge_page_size(1ULL << shift) < 0)
620 * if we have pdshift and shift value same, we don't
621 * use pgt cache for hugepd.
623 if (pdshift > shift) {
624 if (!IS_ENABLED(CONFIG_PPC_8xx))
625 pgtable_cache_add(pdshift - shift);
626 } else if (IS_ENABLED(CONFIG_PPC_E500) ||
627 IS_ENABLED(CONFIG_PPC_8xx)) {
628 pgtable_cache_add(PTE_T_ORDER);
635 pr_info("Failed to initialize. Disabling HugeTLB");
640 arch_initcall(hugetlbpage_init);
642 void __init gigantic_hugetlb_cma_reserve(void)
644 unsigned long order = 0;
647 order = PUD_SHIFT - PAGE_SHIFT;
648 else if (!firmware_has_feature(FW_FEATURE_LPAR) && mmu_psize_defs[MMU_PAGE_16G].shift)
650 * For pseries we do use ibm,expected#pages for reserving 16G pages.
652 order = mmu_psize_to_shift(MMU_PAGE_16G) - PAGE_SHIFT;
655 VM_WARN_ON(order < MAX_ORDER);
656 hugetlb_cma_reserve(order);